Research map / Image generation / Super-resolution
Uncertainty-guided Perturbation for Image Super-Resolution Diffusion Model
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes UPSR, a pixel-space diffusion super-resolution model that uses an uncertainty estimate (residual of an auxiliary SR network) to reduce noise in flat LR regions, plus SR-conditioning and a VQGAN-free architecture.
Problem
Diffusion-based image super-resolution starts from LR-embedded noise with isotropic noise levels, ignoring that flat and edge/texture regions of the LR image lie at different distances from the HR distribution, leading to suboptimal perceptual quality and high computational cost.
Approach
Treat different LR regions as sitting at different diffusion timesteps: estimate per-pixel uncertainty from the residual between an auxiliary SR prediction and the LR input, and use it to weight the noise level (Uncertainty-guided Noise Weighting), reducing noise in flat areas; additionally condition the denoiser on the SR estimate and replace VQGAN with PixelUnshuffle/nearest upsampling for pixel
Key results
- On RealSR, UPSR-5 reaches MUSIQ 64.541 and MANIQA 0.3828 versus 57.769 and 0.3691 for ResShift-15, while using about 30% fewer parameters than ResShift.
- On ImageNet-Test, UPSR-5 improves MUSIQ by 7.21 over ResShift-4 (59.227 vs 52.019) and NIQE by 2.10.
- Training speed increases by 167% (0.45 vs 1.20 s/iter) and GPU memory drops 38% (14.9G vs 24.1G) relative to ResShift.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add uncertainty-guided noise weighting (UNW) (ablation) | ResShift-based SR pipeline · RealSR | CLIPIQA: 0.5447 -> 0.5760 | Helped |
| Add uncertainty-guided noise weighting (UNW) (ablation) | ResShift-based SR pipeline · RealSet | CLIPIQA: 0.6141 -> 0.6340 | Helped |
| Add SR prediction g(y0) as conditional information alongside the LR input (ablation) | ResShift-based SR pipeline with UNW · RealSR | PSNR: +0.34 dB | Helped |
| Add SR prediction g(y0) as conditional information alongside the LR input (ablation) | ResShift-based SR pipeline with UNW · RealSR | CLIPIQA: +0.0247 | Helped |
| Replace VQGAN encoder and decoder with PixelUnshuffle and nearest-neighbor upsampling | UPSR · RealSR | MUSIQ: 48.698 -> 64.541 | Helped |
| Replace VQGAN encoder and decoder with PixelUnshuffle and nearest-neighbor upsampling | UPSR · Training | training speed: increasing training speed by 167% | Helped |
| Replace VQGAN encoder and decoder with PixelUnshuffle and nearest-neighbor upsampling | UPSR · Training | memory footprint: saving GPU memory footprint by 38% | Helped |
Surprising findings
- Reducing noise in flat regions improves perceptual quality rather than degrading it, counter to the expectation that more noise helps generation.
Limitations
- Uncertainty is approximated by the residual of an auxiliary SR network rather than a learned uncertainty estimate, which is a heuristic.
- Gains over ResShift are mostly in perceptual metrics; PSNR and SSIM are not consistently improved (e.g., UPSR-5 PSNR is lower than ResShift-15 on RealSR).
Experimental setup
Datasets: ImageNet ImageNet-Test RealSR RealSet65
Benchmarks: ImageNet-Test RealSR RealSet65
Metrics: PSNR SSIM LPIPS CLIPIQA MUSIQ MANIQA NIQE
Baselines: ESRGAN RealSR-JPEG BSRGAN RealESRGAN SwinIR DASR LDM-15 ResShift-15 ResShift-4
Builds on: ResShift SR3 LDM LPIPS RealESRGAN
Closest papers by meaning
- BUFF: Bayesian Uncertainty Guided Diffusion Probabilistic Model for Single Image Super-Resolution (AAAI 2025 · similarity 0.92)
- Noise-Started One-Step Image Super-Resolution via LR-Conditioned SplitMeanFlow and GAN Refinement (NeurIPS 2026 · similarity 0.91)
- Arbitrary-steps Image Super-resolution via Diffusion Inversion (CVPR 2025 · similarity 0.90)
- SinSR: Diffusion-Based Image Super-Resolution in a Single Step (CVPR 2024 · similarity 0.90)
- One-Step Effective Diffusion Network for Real-World Image Super-Resolution (NeurIPS 2024 · similarity 0.89)
- Unsupervised Diffusion-Based Degradation Modeling for Real-World Super-Resolution (AAAI 2025 · similarity 0.89)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.